Affects Version/s: 6.2.X EE, 7.0.0 M4
Fix Version/s: 7.0.0 M4
The failure is like this:
This is really a jdk bug:
However openjdk guys don't seem to take care of it, as they close it as "Incomplete", saying the error is impossible, but it indeed is possible and happening to us.
The race condition roots from registering new channels during the selector close. Without changing jdk code, it is impossible to fix the root cause.
But we can add one more layer synchronization to prevent registering new channel during the selector close. Theoretically we can avoid this from happening to us.
While our CI constantly running this test, if this does not work, or there is more race condition, soon we will know.
There is nothing for QA to manual test in this ticket.